def epoch_completed_callback(engine):
epoch = engine.state.epoch
if lr_scheduler is not None:
lr_scheduler.step()
if epoch % eval_interval == 0:
logger.info("Model saved at {}/{}_model_{}.pth".format(
save_dir, prefix, epoch))
if evaluator and epoch % eval_interval == 0:
torch.cuda.empty_cache()
# extract query feature
evaluator.run(query_loader)
q_feats = torch.cat(evaluator.state.feat_list, dim=0)
q_ids = torch.cat(evaluator.state.id_list, dim=0).numpy()
q_cams = torch.cat(evaluator.state.cam_list, dim=0).numpy()
# extract gallery feature
evaluator.run(gallery_loader)
g_feats = torch.cat(evaluator.state.feat_list, dim=0)
g_ids = torch.cat(evaluator.state.id_list, dim=0).numpy()
g_cams = torch.cat(evaluator.state.cam_list, dim=0).numpy()
g_img_paths = np.concatenate(evaluator.state.img_path_list, axis=0)
perm = sio.loadmat(
os.path.join(dataset_cfg.sysu.data_root, 'exp',
'rand_perm_cam.mat'))['rand_perm_cam']
mAP, r1, r5, _, _ = eval_sysu(q_feats, q_ids, q_cams, g_feats,
g_ids, g_cams, g_img_paths, perm)
if writer is not None:
writer.add_scalar('eval/mAP', mAP, epoch)
writer.add_scalar('eval/r1', r1, epoch)
writer.add_scalar('eval/r5', r5, epoch)
evaluator.state.feat_list.clear()
evaluator.state.id_list.clear()
evaluator.state.cam_list.clear()
evaluator.state.img_path_list.clear()
del q_feats, q_ids, q_cams, g_feats, g_ids, g_cams
torch.cuda.empty_cache()
logging.basicConfig(level=logging.INFO, format="%(message)s")
# extract feature
cmd = "python{} extract.py {} {} ".format(sys.version[0], args.gpu, args.model_path)
if args.dataset is not None:
cmd += "--dataset {}".format(args.dataset)
subprocess.check_call(cmd.strip().split(" "))
# evaluation
q_mat_path = 'features/%s/query-%s.mat' % (dataset, prefix)
g_mat_path = 'features/%s/gallery-%s.mat' % (dataset, prefix)
assert os.path.exists(q_mat_path)
assert os.path.exists(g_mat_path)
mat = sio.loadmat(q_mat_path)
q_feats = mat["feat"]
q_ids = mat["ids"].squeeze()
q_cam_ids = mat["cam_ids"].squeeze()
mat = sio.loadmat(g_mat_path)
g_feats = mat["feat"]
g_ids = mat["ids"].squeeze()
g_cam_ids = mat["cam_ids"].squeeze()
g_img_paths = mat['img_path'].squeeze()
perm = sio.loadmat(os.path.join(dataset_cfg.get(dataset).data_root, 'exp', 'rand_perm_cam.mat'))['rand_perm_cam']
eval_sysu(q_feats, q_ids, q_cam_ids, g_feats, g_ids, g_cam_ids, g_img_paths, perm)
def train_completed(engine):
torch.cuda.empty_cache()
# extract query feature
evaluator.run(query_loader)
q_feats = torch.cat(evaluator.state.feat_list, dim=0)
q_ids = torch.cat(evaluator.state.id_list, dim=0).numpy()
q_cams = torch.cat(evaluator.state.cam_list, dim=0).numpy()
q_img_paths = np.concatenate(evaluator.state.img_path_list, axis=0)
# extract gallery feature
evaluator.run(gallery_loader)
g_feats = torch.cat(evaluator.state.feat_list, dim=0)
g_ids = torch.cat(evaluator.state.id_list, dim=0).numpy()
g_cams = torch.cat(evaluator.state.cam_list, dim=0).numpy()
g_img_paths = np.concatenate(evaluator.state.img_path_list, axis=0)
# print("best rank1={:.2f}%".format(engine.state.best_rank1))
if dataset == 'sysu':
perm = sio.loadmat(
os.path.join(dataset_cfg.sysu.data_root, 'exp',
'rand_perm_cam.mat'))['rand_perm_cam']
eval_sysu(q_feats,
q_ids,
q_cams,
g_feats,
g_ids,
g_cams,
g_img_paths,
perm,
mode='all',
num_shots=1,
rerank=engine.rerank)
eval_sysu(q_feats,
q_ids,
q_cams,
g_feats,
g_ids,
g_cams,
g_img_paths,
perm,
mode='all',
num_shots=10,
rerank=engine.rerank)
eval_sysu(q_feats,
q_ids,
q_cams,
g_feats,
g_ids,
g_cams,
g_img_paths,
perm,
mode='indoor',
num_shots=1,
rerank=engine.rerank)
eval_sysu(q_feats,
q_ids,
q_cams,
g_feats,
g_ids,
g_cams,
g_img_paths,
perm,
mode='indoor',
num_shots=10,
rerank=engine.rerank)
elif dataset == 'regdb':
print('infrared to visible')
eval_regdb(q_feats,
q_ids,
q_cams,
g_feats,
g_ids,
g_cams,
g_img_paths,
rerank=engine.rerank)
print('visible to infrared')
eval_regdb(g_feats,
g_ids,
g_cams,
q_feats,
q_ids,
q_cams,
q_img_paths,
rerank=engine.rerank)
elif dataset == 'market':
eval_regdb(q_feats,
q_ids,
q_cams,
g_feats,
g_ids,
g_cams,
g_img_paths,
rerank=engine.rerank)
evaluator.state.feat_list.clear()
evaluator.state.id_list.clear()
evaluator.state.cam_list.clear()
evaluator.state.img_path_list.clear()
del q_feats, q_ids, q_cams, g_feats, g_ids, g_cams
torch.cuda.empty_cache()
def epoch_completed_callback(engine):
epoch = engine.state.epoch
if lr_scheduler is not None:
lr_scheduler.step()
if epoch % eval_interval == 0:
logger.info("Model saved at {}/{}_model_{}.pth".format(
save_dir, prefix, epoch))
if evaluator and epoch % eval_interval == 0 and epoch > start_eval:
torch.cuda.empty_cache()
# extract query feature
evaluator.run(query_loader)
q_feats = torch.cat(evaluator.state.feat_list, dim=0)
q_ids = torch.cat(evaluator.state.id_list, dim=0).numpy()
q_cams = torch.cat(evaluator.state.cam_list, dim=0).numpy()
q_img_paths = np.concatenate(evaluator.state.img_path_list, axis=0)
# extract gallery feature
evaluator.run(gallery_loader)
g_feats = torch.cat(evaluator.state.feat_list, dim=0)
g_ids = torch.cat(evaluator.state.id_list, dim=0).numpy()
g_cams = torch.cat(evaluator.state.cam_list, dim=0).numpy()
g_img_paths = np.concatenate(evaluator.state.img_path_list, axis=0)
if dataset == 'sysu':
perm = sio.loadmat(
os.path.join(dataset_cfg.sysu.data_root, 'exp',
'rand_perm_cam.mat'))['rand_perm_cam']
mAP, r1, r5, _, _ = eval_sysu(q_feats,
q_ids,
q_cams,
g_feats,
g_ids,
g_cams,
g_img_paths,
perm,
mode='all',
num_shots=1,
rerank=engine.rerank)
elif dataset == 'regdb':
print('infrared to visible')
mAP, r1, r5, _, _ = eval_regdb(q_feats,
q_ids,
q_cams,
g_feats,
g_ids,
g_cams,
g_img_paths,
rerank=engine.rerank)
print('visible to infrared')
mAP, r1_, r5, _, _ = eval_regdb(g_feats,
g_ids,
g_cams,
q_feats,
q_ids,
q_cams,
q_img_paths,
rerank=engine.rerank)
r1 = (r1 + r1_) / 2
elif dataset == 'market':
mAP, r1, r5, _, _ = eval_regdb(q_feats,
q_ids,
q_cams,
g_feats,
g_ids,
g_cams,
g_img_paths,
rerank=engine.rerank)
if r1 > engine.state.best_rank1:
engine.state.best_rank1 = r1
torch.save(model.state_dict(),
"{}/model_best.pth".format(save_dir))
if writer is not None:
writer.add_scalar('eval/mAP', mAP, epoch)
writer.add_scalar('eval/r1', r1, epoch)
writer.add_scalar('eval/r5', r5, epoch)
evaluator.state.feat_list.clear()
evaluator.state.id_list.clear()
evaluator.state.cam_list.clear()
evaluator.state.img_path_list.clear()
del q_feats, q_ids, q_cams, g_feats, g_ids, g_cams
torch.cuda.empty_cache()